Structural Protein–Ligand Interaction Fingerprints ... - ACS Publications

Aug 13, 2014 - Pipeline Pilot software.17 ECFP retains information about all atom/ bond types and uses one unique integer .... realistic for a typical...
0 downloads 11 Views 2MB Size
This is an open access article published under an ACS AuthorChoice License, which permits copying and redistribution of the article or any adaptations for non-commercial purposes.

Article pubs.acs.org/jcim

Structural Protein−Ligand Interaction Fingerprints (SPLIF) for Structure-Based Virtual Screening: Method and Benchmark Study C. Da and D. Kireev* Center for Integrative Chemical Biology and Drug Discovery, Eshelman School of Pharmacy, University of North Carolina at Chapel Hill, Campus Box 7363, Chapel Hill, North Carolina 27599-7363, United States S Supporting Information *

ABSTRACT: Accurate and affordable assessment of ligand−protein affinity for structure-based virtual screening (SB-VS) is a standing challenge. Hence, empirical postdocking filters making use of various types of structure−activity information may prove useful. Here, we introduce one such filter based upon three-dimensional structural protein−ligand interaction fingerprints (SPLIF). SPLIF permits quantitative assessment of whether a docking pose interacts with the protein target similarly to a known ligand and rescues active compounds penalized by poor initial docking scores. An extensive benchmark study on 10 diverse data sets selected from the DUD-E database has been performed in order to evaluate the absolute and relative efficiency of this method. SPLIF demonstrated an overall better performance than relevant standard methods.



implemented in the MOE software suite.14 Although the SIFt-like approaches proved to be useful postdocking analysis techniques, they also have a number of intrinsic limitations. For instance, inferring bond types from quite imperfect binding poses, may lead to frequent bond-type detection mistakes. Moreover, the bond-type categories (e.g., hydrogen bond, polar, nonpolar, and contact10) used in interaction fingerprints do not account for multiple interaction types, such as cation-π, which would be labeled as merely a contact. Here we introduce a new approach termed structural protein−ligand interaction fingerprints (SPLIF) that also exploits the general idea of quantifying and comparing ligand−protein interactions but does it in a very different way. Particularly, in SPLIF, three-dimensional structures of interacting ligand and protein fragments are explicitly encoded in the fingerprint. Consequently, all possible interaction types that may occur between the fragments (e.g., π−π, CH−π, etc.) are implicitly encoded into SPLIF. The reported fingerprints are wrapped into a normalized quantitative score that expresses the similarity between the interaction profile of a docking pose and that of a reference protein−ligand complex. In order to quantitatively assess the performance of this new approach, we submitted it to a comparative test using it as a postdocking score against a panel of 10 diverse protein targets. The targets along with the sets of respective actives and decoys were selected from the Database of Useful Decoys: Enhanced (DUD-E).15 The purpose of this evaluation was to ascertain if

INTRODUCTION In structure-based virtual screening (SB-VS), each screened compound is submitted to a two-step process. In the first step, a compound is docked to the putative binding pocket of the protein in a number of energetically acceptable binding modes called poses.1 In the second step, the free energy of binding is assessed for each pose by a scoring function.2 While there is now a general consensus that most of the popular docking algorithms perform fairly well in generating sound poses, scoring functions most often fail to adequately evaluate the binding affinity.3−9 As a result, even the optimiztic success rates that are generally reported in SB-VS benchmark studies8,9 might often be insufficient for true ligand discovery when screening large chemical libraries against a novel target with an objective to experimentally test 50−100 virtual hits. Therefore, all possible means must be employed to improve the odds of obtaining a sizable number of confirmed actives from a small set of designated virtual hits. Scoring approaches that can take advantage of known ligand-bound protein structures (e.g., enzyme-bound substrates) are of special interest. In 2003, Deng et al. introduced structural interaction fingerprints (SIFt),10 with an objective to represent and analyze three-dimensional protein−ligand binding interactions by encoding them into a one-dimensional binary string. Construction of SIFt is a twostep process consisting of (i) identification of residues interacting with the ligand and (ii) classification of ligand− residue interactions into any of seven predetermined types (e.g., whether the protein backbone or side-chain is involved, residue acts as an H-bond donor or acceptor, etc.). Later, similar techniques were proposed by Mpamhanga et al.,11 Pérez-Nueno et al.,12 and Marcou and Rognan13 and © 2014 American Chemical Society

Received: May 29, 2014 Published: August 13, 2014 2555

dx.doi.org/10.1021/ci500319f | J. Chem. Inf. Model. 2014, 54, 2555−2561

Journal of Chemical Information and Modeling

Article

Table 1. Targets for SPLIF Benchmarking Collected from DUD-E class kinase

target FAK1 AKT1

protease

ACE

TRYB1 HMDH GPCR

ADRB1

nuclear receptor

MCR PRGR

ion channel

GRIK1

synthase

PGH2

decoysa

Gscore cutoffb

description

PDB

activesa

focal adhesion kinase 1 serine/threonineprotein kinase AKT angiotensinconverting enzyme tryptase beta-1 HMG-CoA reductase Beta-1 adrenergic receptor mineralocorticoid receptor progesterone receptor glutamate receptor ionotropic kainate 1 cyclooxygenase-2

3bz3

100 (71)

5350 (2131)

−6.0

9 (1mp8, 2etm, 2ijm, 3bz3, 4gu6, 4gu9, 4i4e, 4k8a, 4kab)

3cqw

293 (199)

16450 (6131)

−5.0

12 (3cqu, 3cqw, 3mv5, 3mvh, 3ocb, 3ow4, 3qkk, 3cql, 3qkm, 4ekk, 4ekl, 4gv1)

3bkl

282 (277)

16900 (16454)

−2.5

13 (1o86, 1uze, 1uzf, 2c6n, 2oc2, 2xy9, 2xyd, 2ydm, 3bkk, 3bkl, 3l3n, 3nxq, 4bxk)

2zec 3ccw

148 (59) 170 (170)

7650 (1657) 8750 (8456)

−6.0 −2.5

2vt4

247 (240)

15842 (13932)

−4.0

2aa2

94 (66)

5150 (2481)

−6.0

3kba

293 (222)

15650 (12914)

−5.0

1vso

101 (96)

6550 (5980)

−2.5

5 (2f9p, 2f9n, 2zeb, 2zec, 4a6l) 22 (1dq8, 1dq9, 1dqa, 1hw8, 1hw9, 1hwi, 1hwj, 1hwk, 1hwl, 2q1l, 2q6b, 2q6c, 2r4f, 3bgl, 3cct, 3ccw, 3ccz, 3cd0, 3cd5, 3cd7, 3cda, 3cdb) 14 (2vt4, 2y00, 2y01, 2y02, 2y03, 2y04, 2ycw, 2ycx, 2ycy, 2ycz, 3zpq, 3zpr, 4ami, 4amj) 13 (2aa2, 1y9r, 1ya3, 2a3i, 2aa5, 2aa6, 2aa7, 2aax, 2ab2, 2abi, 2oax, 3vhu, 3vhv) 17 (1a28, 1e3k, 1sqn, 1sr7, 1zuc, 2ovh, 2ovm, 2w8y, 3d90, 3g8o, 3hq5, 3kba, 3zr7, 3zra, 3zrb, 4a2j, 4apu) 17 (1txf, 1vso, 1ycj, 2f34, 2f35, 2f36, 2pbw, 2qs1, 2qs2, 2qs3, 2qs4, 2wky, 3gba, 3gbb, 3s2v, 4dld, 4e0x)

3ln1

435 (374)

23150 (17948)

−5.0

refs

27 (1cvu, 1cx2, 1ddx, 1pxx, 3hs5, 3hs6, 3hs7, 3krk, 3ln0, 3ln1, 3mdl, 3nt1, 3ntb, 3ntg, 3olt, 3olu, 3pgh, 3q7d, 3qh0, 3qmq, 3rr3, 3tzi, 4cox, 4e1g, 4fm5, 4llz, 6cox)

a Initial numbers of actives and decoys from DUD-E with the numbers after the Gscore filter included in parentheses. bThe Gscore cutoffs are set to allow all reference ligands to be retained (in hope to retain the most of actives in the test set as well).

SPLIF can outperform and/or bring complementary actives compared to standard or analogous approaches.



MATERIALS AND METHODS

Targets, Ligand Data Sets, and Reference Ligands. The Database of Useful Decoys: Enhanced (DUD-E), a standard test set for virtual screening, was used to validate our fingerprint approach. Ten diverse targets covering six protein classes were randomly selected for this benchmark study (Table 1). The protein structures selected were used as targets for docking of actives and decoys; the resulting poses were processed for subsequent SPLIF generation. All available cocrystallized ligands were retrieved from the Protein Data Bank and used to generate reference SPLIF. In order to assess the diversity of the actives and decoys, we calculated pairwise Tanimoto similarities for all ligands used in this study. The histograms in Figure S1 (see the Supporting Information) show similarity probability distributions for each combination of target/activity-category (active or decoy). In all histograms, mostly low similarity values are well-populated, meaning that, similar to typical virtual screening libraries, the sets are overall diverse but have clusters of closely related compounds. The SD files for all data sets including Gscore, SPLIF, 2D similarity and PLIF scores can be obtained on request from the authors. Docking. Ligands were docked into the active site of the target protein using the Glide program16 in standard docking precision (Glide SP). The binding region was defined by a 20 Å × 20 Å × 20 Å box centered on the reference ligand selected from DUD-E. Default settings were adopted for all the remaining parameters. The top 30 poses were generated for each ligand. Structural Protein−Ligand Interaction Fingerprints (SPLIF). Building the Reference Fingerprint. SPLIF-based rescoring consists of calculating SPLIF for each docking pose and comparing it to that of a reference (e.g., experimentally solved) ligand-protein complex. The essential steps of the algorithm for building a reference SPLIF are depicted in Figure 1. In the first step, a reference protein-bound ligand is inspected for protein−ligand contacts. Two atoms are considered being in a contact if the distance between them is within a specified threshold (4.5 Å in this study). In the second step, for each ligand− protein atom pair, the respective ligand and protein atoms are expanded to circular fragments, i.e., fragments that include the atoms in question and their successive neighborhoods up to a certain

Figure 1. Essential steps of building a reference SPLIF.

distance. In Figure 1, the contacting ligand and protein atoms are enclosed in small dotted circles and the respective circular fragments are enclosed in larger concentric circles. Each type of circular fragment is assigned an identifier. Here, we made use of Extended Connectivity Fingerprints up to the first closest neighbor (ECFP2) as defined in the Pipeline Pilot software.17 ECFP retains information about all atom/ bond types and uses one unique integer identifier to represent one substructure (i.e., circular fragment). In the third step, 3D coordinates are retrieved for all atoms involved in ligand and protein fragments. The major difference of SPLIF from earlierSIFt-typefingerprints is that in SPLIF the interactions are encoded implicitly, as a result of explicitly encoding ligand and protein fragments, whereas in SIFt-like methods the interaction types need to be encoded explicitly, by means of empirical rules. Consequently, most of current SIFt-like implementations handle only a small number of interaction types. By contrast, SPLIF implicitly accounts for all types of local interactions. For example, two parallel aromatic fragments would imply a π−π interaction; a cation fragment positioned on the axis perpendicular to the aromatic plane of another fragment would imply a cation-π interaction and so forth. For SIFt-like fingerprints, if these two types of interaction are not encoded, they will certainly not be identified. While SPLIF records any contacting fragments from the ligand and the protein, meaning the two aromatic fragments and the cation and the aromatic plane here. If the test ligand presents the same fragments, a 2556

dx.doi.org/10.1021/ci500319f | J. Chem. Inf. Model. 2014, 54, 2555−2561

Journal of Chemical Information and Modeling

Article

match to the reference would be assigned by SPLIF, but not by SIFtlike fingerprints. Building Fingerprints for Docking Poses of Test Compounds. The interaction fingerprints for the docking poses of a test compound (an active or a decoy from DUD-E) are computed in a similar fashion to that of the reference fingerprints. SPLIF-Based Similarity. The calculation of a SPLIF-based similarity score is depicted in Figure 2. In the first step, the ECFP identifiers of a

hydrogen bonds, ionic interactions, and surface contacts according to the residues. We applied it to virtual screening here and compared it to our SPLIF approach. The PLIF descriptors for all protein-bound ligands were generated with the default parameter set in MOE. The PLIF similarity was expressed by means of the Tanimoto similarity coefficient. Performance Metrics. We made use of enrichment factors (EF) and EF plots to quantitatively assess the performance of the VS scores under study. The EF plot represents the enrichment for a specific topscoring percentile of the database as a function of the corresponding percentile Px (in log10 scale):

⎛ As ⎞ ⎛ A t ⎞ EF = ⎜ ⎟/⎜ ⎟ ⎝ A s + Ds ⎠ ⎝ A t + Dt ⎠

where As is the number of active ligands in the selected top scoring percentile of the database; Ds is the number of decoys in the selected top scoring percentile of the database; At is the total number of active ligands in the whole database; and Dt is the total number of decoys in the whole database. Enrichment factor indicates the ability of a virtual screening score to increase the proportion of true positives in a respective percentile relative to an average random selection. The logarithmic scale for the X-axis is used to accentuate the contribution of the lower percentiles to the plot.

Figure 2. Essential steps of SPLIF-scoring the docking poses.



test SPLIF (i.e., the SPLIF of a docking pose to be scored) are compared to ECFP identifiers of the reference fingerprint, which is done to find matching circular fragments between the test ligand and the reference ligand and results in a list of 2D-matching SPLIF-bits. In the second step, 3D coordinates of the matching circular fragments (2D-matching bits) are retrieved and root-mean-square deviations (RMSDs) are calculated in order to assess the 3D overlay. The bits for which RMSDs are within a specific threshold (set to 1 Å in this study) are considered as (fully) matching. Next, all atom lists of all matching bits are fused together and deduplicated to form two consolidated lists: (i) unique matching ligand atoms (UMLA) and (ii) unique matching protein atoms (UMPA). Finally, a SPLIF-based similarity score is calculated as follows: SPLIF‐Sim =

NUMLA NUMPA NULA NUPA

(2)

RESULTS AND DISCUSSION The generic benchmark workflow for each target included the following steps: (i) collecting reference ligands from crystal structures; (ii) Glide-based docking and scoring of all reference ligands and test ligands (actives and decoys from DUD-E); (iii) selecting test-ligand poses with plausible Gscore values (to leave out the most awkward poses); (iv) generating SPLIF for reference and test ligands, and calculate SPLIF-based similarity scores; (iv) calculating alternative scores (ligand-based similarity and first-generation interaction fingerprints PLIF); and (iv) calculating performance metrics. The targets for the benchmark study represented six protein classes including kinases, proteases, synthases, GPCRs, nuclear receptors, and ion channels (Table 1). In step iii, the Gscore cutoffs were set on per-target basis (see Table 1) to let all reference ligands be selected (in hope to retain most of the actives from the test set as well). The Gscore filter removed substantial numbers of decoys, while sacrificing a relatively small number of actives. However, the enrichment produced by step iii alone was still insufficient, which emphasized the importance of further filtering. The enrichment plots for the 10 targets are shown in Figure 3 and can be used for a quick visual assessment of the VS scores employed. Ideal EF plots (rendered as black solid lines) produced by hypothetic scores that would rank all actives above the decoys are given for reference. As expected, the EFs resulting from the scoring methods under study differed most significantly at relatively small percentages of selected top scoring compounds (